Skill-tree: finding the Claude Code habits you never touch

Anthropics February 2026 AI Fluency Index studied 9,830 Claude conversations and classified them across 11 observable behaviors and three axes: Description, Discernment, Delegation. I read it and immediately wanted to run the same classification on my own sessions — not to get a score, but to see which of the 11 behaviors I never touch at all.

That gap is what skill-tree is built around. The growth quest doesn't pick a behavior you do occasionally and could do more. It picks the one you avoid entirely — the anti-local-optimum move.

Here's why that matters: if you're already fast at iterative prompting and the tool suggests "do more iterative prompting," you've learned nothing. The useful signal is the blank space. Mine was discernment behaviors — I was delegating constantly, describing well, but almost never explicitly evaluating or pushing back on Claude's output. I didn't know that until I looked at the distribution.

claude plugin marketplace add robertnowell/ai-fluency-skill-cards
claude plugin install skill-tree-ai@ai-fluency-skill-cards

After install, run it in any Claude Code session. The 7-step orchestration — session files → extract messages → remote classifier → archetype assignment → narrative → render → URL — takes 30–60 seconds. You get a stable hosted URL with a tarot-card archetype (seven total, museum art) and the specific behavior flagged as your next session's quest.

The quest persists via a SessionStart hook at ~/.skill-tree/ so it's visible at the top of your next session without re-running the analysis.

Also available as an MCP server (npm install skill-tree-ai) for Cursor, VS Code, and Windsurf.

Live fixture: skill-tree-ai.fly.dev/fixture/illuminator
github.com/robertnowell/skill-tree